Scholarlink

<p>Feature representation diversity serves as a fundamental driver for improving the effectiveness and generalization of unsupervised anomaly detection. In particular, deep random projections, as employed in Deep Isolation Forest (DIF), offer greater partitioning flexibility and enable nonlinear partitioning over subspaces of varying sizes. However, the optimization-free randomness of such projections also injects redundancy, instability, and representational ambiguity that erode representation quality and robustness, thereby constraining detection performance. To address this limitation, this paper proposes the hierarchical feature-selected deep isolation forest (HF-DIF). HF-DIF refines diverse deep random projections via a hierarchical feature-selection scheme, which promotes informative feature diversity while alleviating superfluous features, inconsistency, and representational uncertainty. Specifically, the hierarchical feature-selection scheme is driven by an iForest-based feature evaluator that quantifies feature distinguishability using forest-level aggregates of isolation-tree characteristics, including branch-imbalance and path-length statistics, thereby enabling both local and global refinement of projected representations. Theoretical analysis establishes that HF-DIF has stronger capability than DIF in separating anomalous samples from normal ones. Extensive experiments on 20 benchmark tabular datasets show that HF-DIF consistently outperforms the strongest competing baselines, improving dataset-averaged AUC-ROC by 6.2 percentage points, AUC-PR by 3.5 percentage points, and Precision@K by 5.1 percentage points. Two-sided paired <InlineEquation ID="IEq1"><EquationSource Format="TEX">\(t\)</EquationSource><EquationSource Format="MATHML"><math><mi>t</mi></math></EquationSource></InlineEquation>-tests confirm that HF-DIF significantly outperforms all competing baselines on AUC-ROC and Precision@K, while its AUC-PR advantage remains statistically significant over most baselines. HF-DIF advances unsupervised anomaly detection on tabular data by simultaneously preserving feature diversity and reinforcing discriminability.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

HF-DIF: Hierarchical feature-selected deep isolation forest for unsupervised anomaly detection on tabular data

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月29日

查看全文
  • Xiaoliang Tang,
  • Lulu Chen,
  • Lijuan Zhang

摘要

Feature representation diversity serves as a fundamental driver for improving the effectiveness and generalization of unsupervised anomaly detection. In particular, deep random projections, as employed in Deep Isolation Forest (DIF), offer greater partitioning flexibility and enable nonlinear partitioning over subspaces of varying sizes. However, the optimization-free randomness of such projections also injects redundancy, instability, and representational ambiguity that erode representation quality and robustness, thereby constraining detection performance. To address this limitation, this paper proposes the hierarchical feature-selected deep isolation forest (HF-DIF). HF-DIF refines diverse deep random projections via a hierarchical feature-selection scheme, which promotes informative feature diversity while alleviating superfluous features, inconsistency, and representational uncertainty. Specifically, the hierarchical feature-selection scheme is driven by an iForest-based feature evaluator that quantifies feature distinguishability using forest-level aggregates of isolation-tree characteristics, including branch-imbalance and path-length statistics, thereby enabling both local and global refinement of projected representations. Theoretical analysis establishes that HF-DIF has stronger capability than DIF in separating anomalous samples from normal ones. Extensive experiments on 20 benchmark tabular datasets show that HF-DIF consistently outperforms the strongest competing baselines, improving dataset-averaged AUC-ROC by 6.2 percentage points, AUC-PR by 3.5 percentage points, and Precision@K by 5.1 percentage points. Two-sided paired \(t\)t-tests confirm that HF-DIF significantly outperforms all competing baselines on AUC-ROC and Precision@K, while its AUC-PR advantage remains statistically significant over most baselines. HF-DIF advances unsupervised anomaly detection on tabular data by simultaneously preserving feature diversity and reinforcing discriminability.